GPT-OSS-20B: benchmark results
OpenAI's open-weight GPT-OSS 20B, an Apache-2.0 sparse MoE (~3.6B active) reasoning model (August 2025). Provider: OpenAI. Released 2025-08-05. Access: Open.
Aggregate Rating 1579 ± 4, rank #518 of 1656 rated models, from 2262 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| CIPHER Cross-Record Inference - FIR (Zero-Shot) | 30.2 | Harmonic answer score (%; per question the harmonic mean of | 100 |
| CIPHER Cross-Record Inference - MIMIC (Chain-of-Thought) | 36.4 | Harmonic answer score (%; per question the harmonic mean of | 100 |
| CIPHER Cross-Record Inference - MIMIC (Few-Shot) | 32.3 | Harmonic answer score (%; per question the harmonic mean of | 100 |
| CIPHER Cross-Record Inference - MIMIC (Program-of-Thought) | 30.9 | Harmonic answer score (%; per question the harmonic mean of | 100 |
| GSM-Plus-BN - Critical Thinking (CoT) | 33 | Accuracy (%; critical thinking perturbation, essential infor | 100 |
| GSM-Plus-BN - Digit Expansion | 82.6 | Accuracy (%; digit expansion perturbation, numbers expanded | 100 |
| GSM-Plus-BN - Seed Questions | 96.08 | Accuracy (%; unperturbed seed questions; Bangla grade-school | 100 |
| GeoBenchLLM - GeoQuestions1089 Yes/No | 77 | Accuracy (%) | 100 |
| GeoBenchLLM - NY-POI | 40 | Accuracy (%) | 100 |
| GeoBenchLLM - StepGame | 72 | Accuracy (%) | 100 |
| HELM AIR-Bench 2024 - #34.9: Fake online engagement (fake reviews, fake grassroots support) | 93.33 | Refusal Rate (%) | 100 |
| HELM AIR-Bench 2024 - #34: Mis/disinformation | 96.67 | Refusal Rate (%) | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-oss-20b · How It Works · Data refreshed daily, snapshot 2026-10-20.